Bayesian Model Averaging Naive Bayes (BMA-NB):

Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence
Bayesian Model Averaging Naive Bayes (BMA-NB):
Averaging over an Exponential Number of Feature Models in Linear Time
Ga Wu
Australian National University
Canberra, Australia
wuga214@gmail.com
Scott Sanner
NICTA & Australian National University
Canberra, Australia
ssanner@nicta.com.au
Rodrigo F.S.C. Oliveira
University of Pernambuco
Recife, Brazil
rfsantacruz@gmail.com
to different feature sets, BMA has the advantage over feature
selection that its predictions tend to have lower variance in
comparison to any single model.
While BMA has had some notable successes for averaging over tree-based models such as the context-tree
(CTW) weighting (Willems, Shtarkov, and Tjalkens 1995)
algorithm for sequential prediction, or prunings of decision
trees (Oliver and Dowe 1995), it has not been previously
applied to averaging over all possible feature sets (i.e., the
exponentially sized power set) for classifiers such as NB,
which would correspond to averaging over all nodes of the
feature subset lattice — a problem which defies the simpler
recursions that can be used for BMA on tree structures.
In this paper, we show the somewhat surprising result
that it is indeed possible to exactly evaluate BMA over
the exponentially-sized powerset of NB feature models in
linear-time in the number of features; this yields an algorithm no more expensive to train than a single NB model
with all features, but yet provably converges to the globally
optimal feature subset in the asymptotic limit of data.
This is the first-time we are aware of such an exact lineartime computation result for BMA over the feature subset lattice for any classifier or regressor. While previous work has
proposed algorithms for BMA over the feature subset lattice in generalized linear models such as logistic regression
and linear regression (Raftery 1996; Raftery, Madigan, and
Hoeting 1997; Hoeting et al. 1999), all of these methods either rely on biased approximations or use MCMC methods
to sample the model space with no a priori bound on the
mixing time of such Markov Chains; in contrast our proposal for NB does not approximate and provides an exact
answer in linear-time in the number of features.
We evaluate our novel BMA-NB classifier on a range of
datasets showing that it never underperforms NB (as expected) and sometimes offers performance competitive (or
superior) to classifiers such as SVMs and logistic regression
while taking a fraction of the time to train.
Abstract
Naive Bayes (NB) is well-known to be a simple but effective classifier, especially when combined with feature selection. Unfortunately, feature selection methods
are often greedy and thus cannot guarantee an optimal
feature set is selected. An alternative to feature selection is to use Bayesian model averaging (BMA), which
computes a weighted average over multiple predictors;
when the different predictor models correspond to different feature sets, BMA has the advantage over feature
selection that its predictions tend to have lower variance on average in comparison to any single model.
In this paper, we show for the first time that it is possible to exactly evaluate BMA over the exponentiallysized powerset of NB feature models in linear-time in
the number of features; this yields an algorithm about
as expensive to train as a single NB model with all features, but yet provably converges to the globally optimal
feature subset in the asymptotic limit of data. We evaluate this novel BMA-NB classifier on a range of datasets
showing that it never underperforms NB (as expected)
and sometimes offers performance competitive (or superior) to classifiers such as SVMs and logistic regression while taking a fraction of the time to train.
Introduction
Naive Bayes (NB) is well-known to be a simple but effective
classifier in practice owing to both theoretical justifications
and extensive empirical studies (Langley, Iba, and Thompson 1992; Friedman 1997; Domingos and Pazzani 1997;
Ng and Jordan 2001). At the same time, in domains such
as text classification that tend to have many features, it is
also noted that NB often performs best when combined with
feature selection (Manning, Raghavan, and Schütze 2008;
Fung, Morstatter, and Liu 2011).
Unfortunately, tractable feature selection methods are in
general greedy (Guyon and Elisseeff 2003) and thus cannot
guarantee that an optimal feature set is selected — even for
the training data. An interesting alternative to feature selection is to use Bayesian model averaging (BMA) (Hoeting et
al. 1999), which computes a weighted average over multiple
predictors; when the different predictor models correspond
Preliminaries
We begin by providing a graphical model perspective of
Naive Bayes (NB) classifiers and Bayesian Model Averaging (BMA) that simplify the subsequent presentation and
derivation of our BMA-NB algorithm.
c 2015, Association for the Advancement of Artificial
Copyright Intelligence (www.aaai.org). All rights reserved.
3094
Bayesian Model Averaging (BMA)
yi
xi,k
If we had a class of models m ∈ {1..M }, each of which
specified a different way to generate the feature probabilities P (xi |yi , m), then we can write down a prediction for a
generative1 classification model m as follows:
yi
yi
m
fk
xi
xi,k
k=1..K
k=1..K
i=1..N+1
i=1..N+1
(a)
(b)
P (y, x|m, D) = P (x|m, y, D)P (y|D).
i=1..N+1
While we could select a single model m according to
some predefined criteria, an effective way to combine all
models to produce a lower variance prediction than any single model is to use Bayesian model averaging (Hoeting et
al. 1999) and compute a weighted average over all models:
(c)
Figure 1: Bayesian graphical model representations of (a) Naive
Bayes (NB), (b) Bayesian model averaging (BMA) for generative
models such as NB, and (c) BMA-NB. Shaded nodes are observed.
P (y, x|m, D)
X
=
P (x|m, y, D)P (y|D)P (m|D).
Naive Bayes (NB) Classifiers
Assume we have N i.i.d. data points D = {(yi , xi )} (1 ≤
i ≤ N ) where yi ∈ {1..C} is the discrete class label for datum i and xi = (xi,1 , . . . , xi,K ) is the corresponding vector
of K features for datum i. In a probabilistic setting for classification, given a new feature vector xN +1 (abbreviated x),
our objective is to predict the highest probability class label
yN +1 (abbreviated y) based on our observed data. That is,
for each y, we want to compute
P (y|x, D) ∝ P (y, x, D) =
N
+1
Y
P (yi , xi ),
Here we can view P (m|D) as the weight of model m and
observe that it provides a convex
P combination of model predictions, since by definition: m P (m|D) = 1. BMA has
the convenient property that in the asymptotic limit of data
D, as |D| → ∞, P (m|D) → 1 for the single best model m
(assuming no models are identical) meaning that asymptotically, BMA can be viewed as optimal model selection.
To compute P (m|D), we can apply Bayes rule to obtain
(1)
P (m|D) ∝ P (D|m)P (m) =
where the first proportionality is due to the fact that D and
x are fixed (so the proportionality term is a constant) and
QN +1
the i=1 owes to our i.i.d. assumption for D (we likewise
assume the N +1st data point is also drawn i.i.d.) and the fact
that we can absorb (y, x) = (yN +1 , xN +1 ) into the product
by changing the range of i to include N + 1.
The Naive Bayes independence assumption posits that all
features xi,k (1 ≤ k ≤ K) are conditionally independent
given the class yi . If we additionally assume that P (yi ) and
P (xi,k |yi ) are set to their maximum likelihood (empirical)
estimates given D, we arrive at the following standard result
for NB in (3):
K
Y
i=1
k=1
arg max P (y|x, D) = arg max
y
yN +1
= arg max P (yN +1 )
yN +1
P (yi )
K
Y
N
Y
P (yi , xi |m)
i=1
=
N
Y
P (yi )P (xi |m, yi ). (6)
i=1
Substituting (6) into (5) we arrive at the final simple form,
where we again absorb P (x|m, yN +1 , D)P (y|D) into the
QN +1
renamed N + 1st factor of i=1 :
P (y, x|m, D) ∝
+1
X NY
m
P (yi )P (xi |m, yi ).
(7)
i=1
We pause to make two simple observations on BMA before proceeding to our main result combining BMA and
NB. First, as for NB, we can represent the underlying joint
distribution for (7) as the graphical model in Figure 1(b)
where in this case m is a latent (unobserved) variable that
we marginalize over. Second, unlike the previous result for
NB, we observe that with BMA, we cannot simply absorb
the data likelihood into the proportionality constant since it
depends on the unobserved m and is hence non-constant.
P (xi,k |yi ) (2)
P (xN +1,k |yN +1 )
(5)
m
i=1
N
+1
Y
(4)
(3)
k=1
QN
In (3), we dropped all factors in the i=1 from (2) since
these correspond to data constants in D that do not affect
the arg maxy .
Later when we derive
QN BMA-NB, we’ll see that it is not
possible to drop the i=1 since latent feature selection variables will critically depend on D. To get visual intuition for
this perspective, we pause to provide a graphical model representation of the joint probability for (2) in Figure 1(a) that
we extend later. We indicate all nodes as observed (shaded)
since D and x are given and to evaluate arg maxy , we instantiate y to its possible values in turn.
Main Result: BMA-NB Derivation
Consider a naive combination of the previous sections on
BMA and NB if we consider each model m to correspond to
a different subset of features. Given K features, there are 2K
possible feature subsets in the powerset of features yielding
an exponential number of models m to sum over in (7).
1
A generative model like NB provides a full joint distribution
P (yi , xi ) over class labels and features (Ng and Jordan 2001).
3095
However, if we assume feature selections are independent
of each other (a strong assumption justified in the next section), we can explicitly factorize our representation of m and
exploit this for computational efficiency. Namely, we can
use fk ∈ {0, 1} (1 ≤ k ≤ K) to represent whether feature
xi,k for each datum i is used (1) or not (0).
Considering our model class m to be f = (f1 , . . . , fK ),
we can now combine the naive Bayes factorization of (2)
and BMA from (7) — and their respective graphical models in Figures 1(a,b) — into the BMA-NB model shown
in Figure 1(c) represented by the following joint probability marginalized over the latent model class f and simplified
by exploiting associative and reverse distributive laws:
"K
# N +1
K
X Y
Y
Y
P (y|x, D) =
P (fk )
P (yi ) P (xi,k |fk , yi )
f
k=1
i=1
Feature Model Justification
One may question exactly why implementing P (xi,k |fk , yi )
as in (10) corresponds to a feature selection approach in
Naive Bayes. As already outlined, then fk = 0, this is equivalent to making P (xi,k |fk , yi ) a constant w.r.t. yi and hence
this feature can be ignored when determining the most probable class yi . However there is slightly deeper second justification for this model based on the following feature and
class independence analysis:
QN
QN
i=1 P (xi,k , yi )
i=1 P (xi,k |yi )
=
≥1
(12)
QN
QN
i=1 P (xi,k )P (yi )
i=1 P (xi,k )
On the LHS, we show the ratio of the joint distribution of
feature xi,k and class label yi to the product of their marginal
distributions — a ratio which would equal 1 if the two variables were independent, but which would be greater than 1 if
the variables were dependent — the joint would carry more
information than the product of the marginals. Simply by diQN
viding the LHS through by i=1 (P (yi )/P (yi )) = 1, we
arrive at the equivalent middle term which shows the two
terms used in (10).
From this we can conclude that when feature xi,k and
class label yi are independent (i.e., xi,k carries no information as a predictor of yi ) then P (xi,k |fk , yi ) = P (xi,k ) regardless of fk . Only when P (xi,k |fk , yi ) is predictive would
we expect P (fk = 1|xi,k , yi ) > P (fk = 0|xi,k , yi ) and
hence BMA-NB would tend to include feature k (i.e., models including feature k would have a higher likelihood and
hence a higher model weight), which is precisely the BMA
behavior we desire to weight models with more predictive
features more highly than those with less predictive features.
k=1
" +1
# K
N
+1
Y
XX X NY
Y
...
P (yi )
P (fk )
=
P (xi,k |fk , yi )
f1
f2
fk
i=1
k=1
i=1
"N +1
# K
N
+1
Y
YX
Y
=
P (yi )
P (fk )
P (xi,k |fk , yi )
i=1
k=1 fk
(8)
i=1
Before we proceed further, we need to define specifically what model prior P (fk ) and feature distribution
P (xi,k |fk , yi ) we wish to use for BMA-NB. For P (fk ), we
use the simple unnormalized prior
(
1
if fk = 1
P (fk ) ∝ β
(9)
1 if fk = 0
for some positive constant β.
In a generative model such as NB, what it means for a
feature to be used is quite simple to define:
P (xi,k |yi ) if fk = 1
P (xi,k |fk , yi ) =
(10)
P (xi,k )
if fk = 0
BMA vs. Feature Selection
The analysis in the last subsection suggests that features
which have a higher mutual information (MI) with the class
label will lead to higher weights for models that include
those features. This is encouraging since it is consistent with
the observation that MI is often a good criterion for feature
selection (Manning, Raghavan, and Schütze 2008).
However one may ask: why then bother with BMA-NB
if one could simply use MI as a feature selection criterion
instead? Aside from the tendency of BMA to reduce predictor variance on average compared to the choice of any single model along with the asymptotic convergence of BMA
to the optimal feature subset, there is an additional important reason why BMA might be preferred. Feature selection
requires choosing a threshold on the given criterion (e.g.,
MI), which is another classifier hyperparameter that must be
properly tuned on the training data via cross-validation for
optimal performance. Since this expensive threshold tuning
substantially slows down the training time for naive Bayes
with feature selection, it (somewhat surprisingly) turns out
to be much faster to simply average over all exponential
models in linear time using BMA-NB.
Then whenever fk = 0, P (xi,k |fk , yi ) effectively becomes
a constant that can be ignored in the arg maxy of (2) of NB.
Putting the pieces together and explicitly summing over
fk ∈ {0, 1}, we can continue our derivation from (8):
P (y|x, D) ∝
(11)
"N +1
# K "N +1
#
N +1
Y
Y Y
1 Y
P (xi,k |yi )
P (yi )
P (xi,k ) +
β i=1
i=1
i=1
k=1
And this brings us to a final form for BMA-NB that allows P (y|x, D) to be efficiently computed. Note that here
we have averaged over an exponential number of models f
in linear time in the number of features K. This result critically exploits the fact that NB models do not need to be
retrained for different feature subsets due to the feature conditional independence assumption of NB. As standard for
BMA, we note that in the limit of data D, P (fk |D) → 1 for
the optimal feature selection choice fk = 1 or fk = 0 thus
leading to a linear-time aymptotically optimal feature selection algorithm for NB. Before we proceed to an empirical
evaluation, we pause to discuss a number of design choices
and implementation details that arise in BMA-NB.
Hyperparameter Tuning
Our key hyperparameter to tune in BMA-NB is the constant
β used in the feature inclusion prior (9) for fk . The purpose
3096
of this constant is simply to weight the a priori preference
for models with the feature and without it. Clearly a β = 1
places no prior preference on either model, while β > 1
prefers the simpler model in the absence of any data and
corresponds to a type of Occam’s razor assumption — prefer
the simplest model unless the evidence (Bayesian posterior
for fk ) suggests otherwise.
In experimentation we found that the optimal value of β is
highly sensitive to the amount of data N since the likelihood
QN +1
terms i=1 of (11) scale proportional to N . A stable way
to tune the hyperparameter β across a range of data is to
instead tune Γ defined as
β = ΓN +1
Table 1: UCI datasets for experimentation.
Problem
anneal
autos
vote
sick
crx
mushroom
heart-statlog
segment
labor
vowel
audiology
iris
zoo
lymph
soybean
balance-scale
glass
hepatitis
haberman
(13)
where N is size of training data and we evaluated Γ ∈
{0.5, 0.6, 0.7, 0.8, 0.9, 1, 1.1, 1.2, 1.3, 1.4, 1.5} since, empirically, wider ranges did not lead to better performance.
Conditional Probability Estimation
As usual for NB and other generative classifiers, it is critical to smooth probability estimates for discrete variables to
avoid 0 probabilities (or otherwise low probabilities) for low
frequency features. As often done for NB, we use Dirichlet
prior smoothing for discrete-valued features taking M possible feature values:
V
#D{xk y} + α
P (xk |y) =
(14)
#D{y} + αM
When the Dirichlet prior α = 1, then this approach is called
Laplace smoothing (Mitchell 2010). Note that while α is
also a potential hyperparameter in this work, we keep α = 1
as performance for BMA-NB and NB did not differ significantly for other values of α.
For continuous features xk , we instead model the condi2
tional distribution P (xk |y) = N (µk,y , σk,y
) as a Gaussian
where for each continuous feature k and class label y we
2
estimate the empirical mean µk,y and variance σk,y
for the
corresponding subset of feature data {xi,k |yi = y}i for class
label y.
K
38
26
16
30
16
22
13
19
16
14
69
5
18
19
35
4
10
19
4
|D|
798
205
435
3772
690
8124
270
2310
57
990
226
150
101
148
683
625
214
155
306
C
6
7
2
2
2
2
2
7
2
11
23
3
7
4
19
3
7
2
2
K/|D|
4.76%
12.68%
3.68%
0.80%
2.32%
0.27%
4.81%
0.82%
28.07%
1.41%
30.53%
3.33%
17.82%
12.84%
5.12%
0.64%
4.67%
12.25%
1.31%
Missing?
Yes
Yes
Yes
Yes
No
Yes
No
No
No
No
Yes
No
No
No
Yes
No
No
No
No
Comparison Methodology
In our experimentation, we compare classifiers using a
nested random resampling cross-validation (CV) model to
ensure all classifier hyperparameters were properly tuned via
CV on the training data prior to evaluation on the test data.
All algorithms — NB, BMA, SVM, and Logistic Regression
(LR) — require hyperparameter tuning during nested crossvaliation as outlined below along with other training details
for each algorithm.
The implementation of SVM and LR are from Liblinear
API package (Fan et al. 2008). Both toolkits optimize the
regularized error function
l
X
1
ξ (w; xi ; yi )
min wT w + C
w 2
i=1
(15)
where SVM uses hinge loss for ξ and LR uses log loss for
ξ. Both use a regularization parameter denoted λ, typically
tuned in reciprocal form C = λ1 . C is regarded as penalty
factor or cost factor.2 With large C, the classifiers are more
sensitive to the empirical loss ξ and tend to overfit for high
C, while small C simply prevents learning altogether.
Another parameter of Liblinear package is the constant
bias term B. The following equation corresponds to separating hyperplane for general linear classifiers like SVM and
LR
K
X
y (x, w) =
wk φk (x) + B (w0 )
(16)
Experiments
We evaluated NB, BMA-NB (shortened to just BMA here),
SVM, and Logistic Regression (LR) on a variety of realworld datasets from the UCI machine learning repository (Asuncion and Newman 2007). Characteristics of the
datasets we evaluated are outlined in Table 1.
We compare to the NB classifier with the full feature set
to determine whether BMA always provides better performance than NB as we might expect (since it should place
high weight on the full feature set used by NB if that feature
set offers best performance). We choose both SVM and LR
as discriminative linear classifiers known for their typically
stronger performance but slower training time compared to
NB. While we would not expect generatively trained BMA
to beat disriminatively trained SVM and LR, we are interested to see (a) where BMA falls in the range between SVM
and LR, (b) whether it is significantly faster to train than
SVM and LR, and (c) if there are cases where BMA is competitive with SVM and LR.
k=1
where φk is base function, w is parameter vector and B is
the bias.3
2
We use C to be consistent with the Liblinear software package
we use, but this is not to be confused with the number of classes C;
the intended usage should be clear from context.
3
Note it is not a good idea to regularize B; it is often better to
tune it via cross-validation.
3097
Table 2: (left) Error ±95% confidence interval for each classifier with the best performance average performance shown in bold
— lower is better; (right) training time (ms). Datasets that show a 0.00 error rate appear to be linearly separable.
Problems
anneal
autos
vote
sick
crx
mushroom
heart-statlog
segment
labor
vowel
audiology
iris
zoo
lymph
soybean
balance-scale
glass
hepatitis
haberman
BMA
7.87±2.00
30.00±7.13
4.65±2.25
6.37±0.89
23.19±3.58
1.35±0.29
7.41±3.55
23.81±1.97
0.00±0.00
33.33±3.34
22.73±6.21
6.67±4.54
0.00±f0.00
7.14±4.72
11.76±2.75
9.68±2.63
52.38±7.61
26.67±7.91
33.33±6.00
NB
7.87±2.00
40.00±7.62
11.63±3.42
7.69±0.97
23.19±3.58
3.82±0.47
7.41±3.55
23.81±1.97
0.00±0.00
42.42±3.50
27.27±6.60
13.33±6.18
0.00±0.00
7.14±4.72
11.76±2.75
9.68±2.63
52.38±7.61
26.67±7.91
33.33±6.00
LR
6.74±1.86
25.00±6.74
4.65±2.25
2.65±0.58
18.84±3.32
2.96±0.42
18.52±5.27
10.82±1.44
0.00±0.00
44.44±3.52
9.09±4.26
6.67±4.54
0.00±0.00
14.29±6.41
10.29±2.59
16.13±3.28
33.33±7.18
33.00±8.77
36.67±6.14
SVM
17.98±2.86
55.00±7.74
4.65±2.25
4.51±0.75
21.74±3.50
4.06±0.49
11.11±4.26
14.72±1.64
0.00±0.00
47.47±3.54
13.64±5.09
6.67±4.54
0.00±0.00
14.29±6.41
11.76±2.75
12.90±2.99
47.62±7.61
46.67±8.93
33.33±6.00
Since parameter tuning of SVM and LR is very time consuming, we limited our hyperparameter search to a combination of 36 joint values of B and C as shown below:
C ∈ {0.01, 0.1, 1, 10, 20, 100}
(17)
B ∈ {0.01, 0.1, 1, 10, 20, 100}
(18)
We implement the Naive Bayes classifier ourselves to
guarantee that NB is exactly the special case of BMA where
weight 1 is placed on the model with f = 1 and weight 0 on
other models. As described previously, while β is tuned for
BMA, α = 1 yielded best results for both NB and BMA and
setting α = 1 for both classifiers allows us to evaluate the
performance of BMA’s model averaging vs. the same full
model that NB uses.
Problems
anneal
autos
vote
sick
crx
mushroom
heart-statlog
segment
labor
vowel
audiology
iris
zoo
lymph
soybean
balance-scale
glass
hepatitis
haberman
BMA
526
81
10
202
36
248
18
435
3
105
81
4
6
7
99
15
14
7
4
NB
147
40
6
79
9
98
5
120
3
35
14
2
3
3
18
5
5
3
2
LR
618
183
22
434
61
1072
22
2606
6
557
172
10
31
22
559
40
50
16
11
SVM
2070
1278
33
2475
540
5785
213
6637
27
5266
260
102
222
175
818
542
621
120
157
To continue with our experimentation, we now perform
some specialized experiments for text classification with the
UCI newsgroups dataset to experiment with BMA performance relative to other classifiers as the amount of data
and number of features changes. We choose text classification for this analysis since both data and features are abundant in such tasks and they are also a special case where
NB is known to perform relatively well, especially with
feature selection (Manning, Raghavan, and Schütze 2008;
Fung, Morstatter, and Liu 2011).
We focus on a binary classification problem which is abstracted from the newsgroup dataset. To construct D for our
first task, we chose 1000 data entries selected from two categories of newsgroups — it is an exactly balanced dataset.
The K = 1000 features are selected from 20,000 words with
high frequency of occurrence after removing English stopwords. Results for this text classification analysis are shown
in Figure 2. We average this plot over a large number of
runs to obtain relatively smoothed average performance results, but omit dense confidence intervals to keep the plots
readable.
In Figure 2(left), we observe that after around 150 features selected, BMA, SVM and LR show significantly better
performance than NB. Furthermore, the accuracy of BMA,
SVM and Logistic Regression also asymptote earlier with
BMA showing competitive performance with SVM and LR
for this text classification task with a large number of (correlated) word features.
To show whether training dataset size can significantly
impact the accuracy of news classification task, we refer
to Figure 2(right), which is similar to the setup for Figure 2(left) except with |D| = 4000 and K = 500. These
results show the discriminative classifiers SVM and LR performing better than BMA, but overall asymptoting more
quickly than NB indicating an ability of BMA to place heav-
Experimental Results
Now we proceed to compare each of the classifiers on the
UCI datasets as shown in Table 2.
Here we observe a few general trends:
• BMA never underperforms NB as expected. LR seems to
offer the best performance, edging out SVM since we generally found that LR was more robust during hyperparameter tuning in comparison to SVM.
• In the range of performance between NB and LR/SVM,
BMA twice outperforms LR (when NB performance is
significantly worse) and outperforms SVM even more often! When BMA does worse than LR and SVM, it often
comes close to the best performance in many cases.
• BMA is up to 4 times slower than NB, but it’s significant performance improvement over BN seems worth this
tradeoff.
• BMA is always faster than LR and SVM — up to 5 times
faster than LR in some cases and up to 50 times faster than
the SVM for glass!
3098
Figure 2: (left) Accuracy vs. number of features, (right) accuracy vs. number of training data.
Figure 3: Effect of tuning hyperparameter Γ for BMA on the UCI datasets.
Table 2 where we see that BMA performance matches NB
performance (which uses all features) for the problems in
Figure 3(right).
ier weight on the most useful features when data is limited
and many features are noisy.
Finally we show results for hyperparameter tuning of
BMA across the UCI datasets in Figure 3. We break the results into two sets — Figure 3(left) where there seems to be
an optimal choice for β and Figure 3(right) where the optimal choice seems to be a small β < 1 which places a strong
prior on using all features.
If we examine the dataset characteristics in Table 1, we
see all of the problems in Figure 3(right) have small numbers
of features and large amounts of data; hence it seems that in
these settings all features are useful and thus the hyperparameter tuning is largest for a prior which selects all features.
This is further corroborated by the performance results in
Conversely, we might infer that in Figure 3(left), not all
features in these problems are useful, thus a peak value for
prior β > 1 is best which indicates a preference to exclude
features unless the data likelihood overrides this prior. For
many of these problems in Figure 3(left) we similarly note
that the performance of BMA was better than NB indicating
that BMA may have placed low weight on models with certain (noisy) features, whereas NB was constrained to use all
features and hence performed worse.
3099
Fan, R.-E.; Chang, K.-W.; Hsieh, C.-J.; Wang, X.-R.; and
Lin, C.-J. 2008. Liblinear - a library for large linear classification. The Weka classifier works with version 1.33 of
LIBLINEAR.
Friedman, J. H. 1997. On bias, variance, 0/1 - loss, and
the curse-of-dimensionality. Data Mining and Knowledge
Discovery 1:5577.
Fung, P. C. G.; Morstatter, F.; and Liu, H. 2011. Feature selection strategy in text classification. In Advances in Knowledge Discovery and Data Mining. Springer. 26–37.
Guyon, I., and Elisseeff, A. 2003. An introduction to variable and feature selection. J. Mach. Learn. Res. 3:1157–
1182.
Hoeting, J. A.; Madigan, D.; Raftery, A. E.; and Volinsky,
C. T. 1999. Bayesian model averaging: A tutorial. Statistical
Science 14(4):382–417.
Lafferty, J. D.; McCallum, A.; and Pereira, F. C. N. 2001.
Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the
Eighteenth International Conference on Machine Learning
(ICML-01), ICML ’01, 282–289.
Langley, P.; Iba, W.; and Thompson, K. 1992. An analysis
of bayesian classifiers. In AAAI-92.
Manning, C. D.; Raghavan, P.; and Schütze, H. 2008. Introduction to Information Retrieval. New York, NY, USA:
Cambridge University Press.
Mitchell, T. M. 2010. Generative and discriminative classifier: Naive bayes and logistic regression.
Ng, A. Y., and Jordan, M. I. 2001. On discriminative vs.
generative classifiers: A comparison of logistic regression
and naive bayes. In NIPS, 841–848.
Oliver, J. J., and Dowe, D. L. 1995. On pruning and averaging decision trees. In ICML-95, 430–437. 340 Pine Street,
6th Floor San Francisco, California 94104 U.S.A.: Morgan
Kaufmann.
Raftery, A., and Dean, N. 2006. Variable selection for
model-based clustering. Journal of the American Statistical
Assocation 101:168–178.
Raftery, A.; Karny, M.; and Ettler, P. 2010. Online prediction under model uncertainty via dynamic model averaging:
Application to a cold rolling mill. Technometrics 52:52–66.
Raftery, A. E.; Madigan, D.; and Hoeting, J. A. 1997.
Bayesian model averaging for linear regression models.
Journal of the American Statistical Association
92(437):179–191.
Raftery, A. E. 1996. Approximate Bayes Factors and Accounting for Model Uncertainty in Generalized Linear Models. Biometrika 83(2):251–266.
Willems, F. M. J.; Shtarkov, Y. M.; and Tjalkens, T. J. 1995.
The context-tree weighting method: basic properties. IEEE
Transactions on Information Theory 41(3):653–664.
Conclusion
In this work we presented a novel derivation of BMA for
NB that averaged over the exponential powerset of feature
models in linear time in the number of features. This tends
to yield a lower variance NB predictor (a result of averaging multiple estimators) that converges to the selection of
the best feature subset model in the asymptotic limit of data.
This is the first such result we are aware of for exact BMA
over the powerset of features that does not require approximations or sampling.
Empirically, we observed strong performance from BMANB in comparison to NB, SVMs, and logistic regression
(LR). These results suggest that (1) in return for a fairly constant 4 times slow-down (no matter what dataset size or feature set size), BMA-NB never underperforms NB; (2) sometimes it performs as well as or better than LR or SVM; (3)
it trains faster than LR and SVM and in the best case is up
to 50 times faster than SVM training. These results suggest
that BMA-NB may be a superior general replacement for
NB, and in some cases BMA-NB may even be a reasonable
replacement for LR and SVM, especially when efficient online training is required.
Future work should examine whether the derivational approach used in this paper can extend linear-time BMA over
all feature subsets to other problems. It may be difficult to
do this for the class of discriminative probabilistic classifiers such as logistic regression since each feature subset
leads to a different set of optimal parameters (unlike NB
where the learned parameters for each feature are independent thus preventing the need to retrain NB for each model).
However, BMA has also been applied beyond classification to other problems such as unsupervised learning (Dean
and Raftery 2010; Raftery and Dean 2006) and sequential
models (Raftery, Karny, and Ettler 2010), where techniques
developed in this work may apply. For example, hidden
Markov models (HMMs) are an important generative sequential model with observation independence (as in NB)
that may admit extensions like that developed in BMA-NB;
it would be interesting to explore how such HMM extensions
compare to state-of-the-art linear chain conditional random
fields (CRFs) (Lafferty, McCallum, and Pereira 2001).
Acknowledgements
NICTA is funded by the Australian Government as represented by the Department of Broadband, Communications
and the Digital Economy and the Australian Research Council through the ICT Centre of Excellence program. Rodrigo
F.S.C. Oliveira was sponsored by CNPq - Brazil.
References
Asuncion, A., and Newman, D. J. 2007. UCI machine learning repository.
Dean, N., and Raftery, A. 2010. Latent class analysis variable selection. Annals of the Institute of Statistical Mathematics 62:11–35.
Domingos, P., and Pazzani, M. 1997. On the optimality of
the simple Bayesian classifier under zero-one loss. Machine
Learning 29:103130.
3100